LayerEditor distills edit priors to turn image edits into reusable RGBA layers
jingdong1 · hf · 2026-10-23
LayerEditor unifies instruction-guided image editing with reusable RGBA layer generation: built on an MMDiT with LoRA, it predicts separate background, text, object, and decoration layers while editing, enabling precise local manipulation and layer reuse.
- Method: joint RGB+Alpha token modeling with Compositional Edit-Prior Distillation (C-EPD), a teacher-student scheme aligning direct-edit and layer-composite denoising trajectories.
- Data: LayerEditor-100K dataset of 100k layered triplets spanning text translation/replacement/movement, object add/delete/edit, and whole-poster editing (text edits 48.3%, object edits 36.9%).
- Use case: swap a poster's city photo from Amsterdam to Venice, decompose into six layers, then independently translate the text layer or restyle the background for PSD-like re-editing.
More from Multimodal
- Seed Audio 1.5 preview: one prompt generates dialogue, SFX, ambience and music — FellMentKE · 2026-10-23
- distil_labs post-trains a small model in under 6 hours to build a Phonon-voiced agent — alexcovo_eth · 2026-10-23
- AnyGameRenderer turns AI agent-built graybox game footage into high-fidelity visuals — AlayaLab · 2026-10-23
- NVIDIA's ArtiFixer fixes 3D reconstruction artifacts with 8B model, 3-4x faster than 14B — ZGojcic · 2026-10-23
- Midjourney v8.2 showcase: Addams Family oil-painting style with shared sref code — azed_ai · 2026-10-23
- Magnific releases CHUMBO, issue 2 of the Contanimation AI animation anthology — gen_ericai · 2026-10-23